Return

Rebuilding confidence in community participation in the AI age

Stack Overflow · 2025

Role
Design, research & strategy
Team
PM & tech lead
Timeline
4 months · 2025

Context

Generative AI drew developers away from Stack Overflow, reducing questions, answers and engagement. At the same time, about 40% of revenue came from licensing community knowledge to AI companies, even as the platform produced less of that content.

I focused on activation: turning passive visitors into active contributors to increase the volume of high-quality questions and answers on the platform.

A Stack Overflow question page

The problem

Developers trusted Stack Overflow, but contributing felt intimidating. Strict moderation, high quality standards and the risk of downvotes or deletion meant new users often learned the rules only after making a mistake. As a result, most visitors read without posting.

Impact

What I did

Understanding why developers hesitated to participate

Developers returned to Stack Overflow when AI answers were unreliable, but doubted they would get help quickly.

Two remote interview sessions with developers
Developer interviews.
Async analysis board with eight interviews broken down into participant intro, problem-solving story, AI didn't solve, getting help on Stack Overflow and ideas to improve
Async interview analysis.
Troubleshooting journey: problem discovery, AI troubleshooting, the AI fallback gap, then validated resolution on Stack Overflow, with sentiment dipping when the AI answer falls short
Troubleshooting journey.

Testing two ways to rebuild confidence

We tested Live 1:1 help and Liveliness using fake-door and A/B tests to explore how the platform could become a more reliable fallback.

1:1 Help. A direct line to someone who could help, for moments when public asking felt intimidating.
Liveliness. Surfacing real community activity so new users could see the platform was alive.
The asker test. We collected questions from askers and used them to test the helper side of the experiment.
The helper test. Helpers reviewed questions from askers and indicated whether they could help.

Choosing the lower-risk path

The test showed stronger interest from helpers than askers. Among helpers, 9.9% clicked through and 1.7% offered to help. Among askers, 1.8% clicked through and 0.3% requested support.

Based on the results, I argued against building 1:1 help and shipped Liveliness as a low-risk improvement we could monitor in production.

We then expanded the Liveliness widget to surface a broader range of community activity signals and monitored its impact in production.

Learning

The best decision I made that quarter was choosing not to ship.